EN

Mistral to Open-Source New Flagship Model, Claims Top Non-Chinese Ranking

Sukaina Khalid

Key Points

  1. Mistral opened a public preview Tuesday of Mistral Large 4, with weights due October 27.
  2. The French startup trained the model on roughly 4,000 Nvidia Grace Blackwell GPUs in European data centers.
  3. Mistral is betting open weights can win enterprise customers away from OpenAI and Anthropic.

The latest:

A public preview of Mistral Large 4, nicknamed Le Chonk, went live Tuesday, with the French startup saying it will publish the model’s trained weights on October 27. Mistral said the model would rank among the top open-weight systems globally by aggregate benchmark performance, and would be the strongest open-weight model built outside China by a substantial margin.

Details:

  • The release plan: Mistral said it opened the preview first so developers, cybersecurity specialists and state authorities could evaluate how the model behaves before the weights are published. Open-weight means the trained parameters are downloadable and modifiable by anyone, unlike closed systems where code, training data and weights stay hidden.
  • The hardware: The company said it trained Mistral Large 4 on about 4,000 Nvidia Grace Blackwell graphics processing units deployed in its own European data centers. Mistral did not disclose the cost of the training run or the length of time it took.
  • The claim: Mistral said the model would place among the best open-weight models worldwide on aggregate benchmark performance. It framed the comparison in two tiers: global open-weight standing, and a clear lead over every open-weight model developed outside China.
  • The security argument: Mistral said closed models can pose security risks for users because they sometimes refuse to answer legitimate cyberdefense requests. The company has argued since its founding that open weights let customers tailor systems to their own needs while protecting their data and intellectual property.
  • The executive line: Guillaume Lample, Mistral co-founder and chief scientist, said the model’s capabilities “will further improve as we scale up our training capacity.” He did not give a target for that expanded capacity or a timeline for the next model generation.
  • The money: The announcement lands roughly a month after Mistral closed a funding round led by South Korea’s Samsung Electronics that pushed its valuation above $24 billion. The company said the raise gives it additional firepower to invest in model development.
  • The competition: Mistral framed the launch as part of a push to widen its appeal, lock in more customers and compete with U.S. rivals, naming ChatGPT maker OpenAI and Anthropic. Both sell primarily closed models, making the weights question a commercial dividing line, not only a technical one.
  • The disclosure: News Corp, which owns Dow Jones Newswires and The Wall Street Journal, has a content-licensing partnership with OpenAI, one of the rivals Mistral named.

Background:

Mistral has positioned itself since launch as Europe’s open-weight counterweight to U.S. labs, publishing downloadable models rather than restricting access through paid interfaces. The Samsung-led round that lifted its valuation past $24 billion was its largest to date.

Between the lines:

The carve-out in Mistral’s own claim is telling: it leads on open-weight models outside China, which concedes that Chinese labs set the pace in the category Mistral has built its identity around. The security framing is also a sales argument — Mistral’s cyberdefense critique of closed models targets precisely the enterprise and government buyers OpenAI and Anthropic serve with hidden weights.

What’s next

October 27 is the date to watch, when Mistral says it will publish the weights and third parties can test the benchmark claims independently. Also watch whether Samsung’s investment translates into additional training capacity Lample referenced.

What to read next